基于域知识迁移的绝缘子缺陷检测方法

绝缘子缺陷检测模型在应用过程中，需要对多种目标域图像进行检测。然而，由于不同目标域图像的分布不同，会导致在源域训练得到的缺陷检测模型难以在不同目标域中实现较好的检测效果。针对缺陷检测模型泛化性较低的问题，提出了基于域知识迁移的绝缘子缺陷检测方法。首先，采用一种新的特征融合方法提高对缺陷目标的特征提取能力；然后通过对抗性训练实现基于图像和实例级别的域分类器，在图像和目标层级上实现了域迁移；最后在区域建议网络中引入区域级别的域分类器，进一步实现了源域和目标域在区域级别上的对齐。实验结果表明，模型能够缓解不同域图像分布不平衡导致检测效果差的问题，并且与其他多种目标检测方法相比，实现了更好的缺陷检测效果。

Research on Insulator Defect Detection Method Based on Domain Knowledge Transfer

In the application process of insulator defect detection model, it is necessary to detect a variety of target domain images. However, due to the different distribution of images in different target domains, it is difficult for the defect detection model trained in the source domain to achieve good detection results in different target domains. Aiming at the problem of low generalization of defect detection models, a transmission line defect detection method based on domain knowledge transfer was proposed. First, a new feature fusion method was adopted to im-prove the feature extraction ability of defective targets; then image- and instance-level-based domain classifiers were implemented through adversarial training, and domain transfer was realized at image and target levels; A region-level domain classifier was introduced into the network, which further realized the alignment of the source and target domains at the region level. The experimental results show that the model can alleviate the problem of poor detection effect caused by the imbalance of image distribution in different domains, and achieve better defect detection effect compared with other target detection methods.